<s>
Zero-shot	B-Algorithm
learning	I-Algorithm
(	O
ZSL	O
)	O
is	O
a	O
problem	O
setup	O
in	O
machine	O
learning	O
where	O
,	O
at	O
test	O
time	O
,	O
a	O
learner	O
observes	O
samples	O
from	O
classes	O
which	O
were	O
not	O
observed	O
during	O
training	O
,	O
and	O
needs	O
to	O
predict	O
the	O
class	O
that	O
they	O
belong	O
to	O
.	O
</s>
<s>
Zero-shot	B-Algorithm
methods	O
generally	O
work	O
by	O
associating	O
observed	O
and	O
non-observed	O
classes	O
through	O
some	O
form	O
of	O
auxiliary	O
information	O
,	O
which	O
encodes	O
observable	O
distinguishing	O
properties	O
of	O
objects	O
.	O
</s>
<s>
This	O
problem	O
is	O
widely	O
studied	O
in	O
computer	B-Application
vision	I-Application
,	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
and	O
machine	B-General_Concept
perception	I-General_Concept
.	O
</s>
<s>
The	O
first	O
paper	O
on	O
zero-shot	B-Algorithm
learning	I-Algorithm
in	O
natural	B-Language
language	I-Language
processing	I-Language
appeared	O
in	O
2008	O
at	O
the	O
AAAI’08	O
,	O
but	O
the	O
name	O
given	O
to	O
the	O
learning	O
paradigm	O
there	O
was	O
dataless	O
classification	O
.	O
</s>
<s>
The	O
first	O
paper	O
on	O
zero-shot	B-Algorithm
learning	I-Algorithm
in	O
computer	B-Application
vision	I-Application
appeared	O
at	O
the	O
same	O
conference	O
,	O
under	O
the	O
name	O
zero-data	O
learning	O
.	O
</s>
<s>
The	O
term	O
zero-shot	B-Algorithm
learning	I-Algorithm
itself	O
first	O
appeared	O
in	O
the	O
literature	O
in	O
a	O
2009	O
paper	O
from	O
Palatucci	O
,	O
Hinton	O
,	O
Pomerleau	O
,	O
and	O
Mitchell	O
at	O
NIPS’09	O
.	O
</s>
<s>
This	O
direction	O
was	O
popularized	O
later	O
in	O
another	O
computer	B-Application
vision	I-Application
paper	O
and	O
the	O
term	O
zero-shot	B-Algorithm
learning	I-Algorithm
caught	O
up	O
,	O
as	O
a	O
take-off	O
on	O
one-shot	O
learning	O
that	O
was	O
introduced	O
in	O
computer	B-Application
vision	I-Application
years	O
earlier	O
.	O
</s>
<s>
In	O
computer	B-Application
vision	I-Application
,	O
zero-shot	B-Algorithm
learning	I-Algorithm
models	O
learned	O
parameters	O
for	O
seen	O
classes	O
along	O
with	O
their	O
class	O
representations	O
and	O
rely	O
on	O
representational	O
similarity	O
among	O
class	O
labels	O
so	O
that	O
,	O
during	O
inference	O
,	O
instances	O
can	O
be	O
classified	O
into	O
new	O
classes	O
.	O
</s>
<s>
In	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
the	O
key	O
technical	O
direction	O
developed	O
builds	O
on	O
the	O
ability	O
to	O
"	O
understand	O
the	O
labels	O
"	O
—	O
represent	O
the	O
labels	O
in	O
the	O
same	O
semantic	O
space	O
as	O
that	O
of	O
the	O
documents	O
to	O
be	O
classified	O
.	O
</s>
<s>
This	O
supports	O
the	O
classification	O
of	O
a	O
single	O
example	O
without	O
observing	O
any	O
annotated	O
data	O
,	O
the	O
purest	O
form	O
of	O
zero-shot	B-Algorithm
classification	O
.	O
</s>
<s>
The	O
original	O
paper	O
made	O
use	O
of	O
the	O
Explicit	B-General_Concept
Semantic	I-General_Concept
Analysis	I-General_Concept
(	O
ESA	O
)	O
representation	O
but	O
later	O
papers	O
made	O
use	O
of	O
other	O
representations	O
,	O
including	O
dense	O
representations	O
.	O
</s>
<s>
The	O
original	O
paper	O
also	O
points	O
out	O
that	O
,	O
beyond	O
the	O
ability	O
to	O
classify	O
a	O
single	O
example	O
,	O
when	O
a	O
collection	O
of	O
examples	O
is	O
given	O
,	O
with	O
the	O
assumption	O
that	O
they	O
come	O
from	O
the	O
same	O
distribution	O
,	O
it	O
is	O
possible	O
to	O
bootstrap	O
the	O
performance	O
in	O
a	O
semi-supervised	O
like	O
manner	O
(	O
or	O
transductive	B-General_Concept
learning	I-General_Concept
)	O
.	O
</s>
<s>
Unlike	O
standard	O
generalization	B-Algorithm
in	O
machine	O
learning	O
,	O
where	O
classifiers	O
are	O
expected	O
to	O
correctly	O
classify	O
new	O
samples	O
to	O
classes	O
they	O
have	O
already	O
observed	O
during	O
training	O
,	O
in	O
ZSL	O
,	O
no	O
samples	O
from	O
the	O
classes	O
have	O
been	O
given	O
during	O
training	O
the	O
classifier	O
.	O
</s>
<s>
It	O
can	O
therefore	O
be	O
viewed	O
as	O
an	O
extreme	O
case	O
of	O
domain	B-General_Concept
adaptation	I-General_Concept
.	O
</s>
<s>
Naturally	O
,	O
some	O
form	O
of	O
auxiliary	O
information	O
has	O
to	O
be	O
given	O
about	O
these	O
zero-shot	B-Algorithm
classes	O
,	O
and	O
this	O
type	O
of	O
information	O
can	O
be	O
of	O
several	O
types	O
.	O
</s>
<s>
While	O
this	O
approach	O
was	O
used	O
mostly	O
in	O
computer	B-Application
vision	I-Application
,	O
there	O
are	O
some	O
examples	O
for	O
it	O
also	O
in	O
natural	B-Language
language	I-Language
processing	I-Language
.	O
</s>
<s>
As	O
pointed	O
out	O
above	O
,	O
this	O
has	O
been	O
the	O
key	O
direction	O
pursued	O
in	O
natural	B-Language
language	I-Language
processing	I-Language
.	O
</s>
<s>
a	O
zero-shot	B-Algorithm
classifier	O
can	O
predict	O
that	O
a	O
sample	O
corresponds	O
to	O
some	O
position	O
in	O
that	O
space	O
,	O
and	O
the	O
nearest	O
embedded	O
class	O
is	O
used	O
as	O
a	O
predicted	O
class	O
,	O
even	O
if	O
no	O
such	O
samples	O
were	O
observed	O
during	O
training	O
.	O
</s>
<s>
The	O
above	O
ZSL	O
setup	O
assumes	O
that	O
at	O
test	O
time	O
,	O
only	O
zero-shot	B-Algorithm
samples	O
are	O
given	O
,	O
namely	O
,	O
samples	O
from	O
new	O
unseen	O
classes	O
.	O
</s>
<s>
In	O
generalized	O
zero-shot	B-Algorithm
learning	I-Algorithm
,	O
samples	O
from	O
both	O
new	O
and	O
known	O
classes	O
,	O
may	O
appear	O
at	O
test	O
time	O
.	O
</s>
<s>
Zero	B-Algorithm
shot	I-Algorithm
learning	O
has	O
been	O
applied	O
to	O
the	O
following	O
fields	O
:	O
</s>
